Deep-Learning-Based Detection of Infants with Autism Spectrum Disorder Using Auto-Encoder Feature Representation

Area of Science:

• Neuroscience
• Developmental Psychology
• Speech-Language Pathology

Background:

• Autism spectrum disorder (ASD) diagnosis relies on subjective clinical assessments, posing challenges in accuracy and efficiency.
• Automated methods using vocal characteristics show promise for objective ASD detection, but data limitations hinder widespread application.
• Existing deep learning models for ASD detection struggle with unrefined, diverse datasets and data anonymity concerns.

Discussion:

• This research presents a pre-trained feature extraction auto-encoder model and a joint optimization scheme to enhance ASD detection from speech data.
• The proposed method achieves robustness with widely distributed and unrefined data, overcoming common challenges in autism diagnostics.
• The auto-encoder model effectively extracts relevant speech features, improving the performance of deep learning models for ASD detection.

Key Insights:

• The auto-encoder-based feature extraction significantly improves the performance of ASD detection in infants compared to using raw speech data.
• The joint optimization scheme enhances the model's ability to handle unrefined and diverse datasets, crucial for real-world diagnostic applications.
• This approach offers a more objective, efficient, and potentially more accessible method for early ASD identification.

Outlook:

• Further validation of this deep learning approach across diverse infant populations is essential.
• Integration of this method into clinical settings could revolutionize early ASD screening and diagnosis.
• Future research may explore incorporating other multi-modal data alongside vocal biomarkers for even more precise ASD detection.